This article is very thorough, but makes one fairly big mistake: it assumes there is low variability in task size.
This could be a reasonable assumption if you're very good at adjusting scope, but I would be skeptical unless you have measured it and have the numbers to prove that in practise, tasks indeed turn out to be roughly the same size (±20 % or whatever -- I don't know the exact threshold.)
Why this matters is that under cheap context switches and even modest task size variability, it's -- surprisingly -- a performant queuing policy to pre-empt the currently processing task to handle every incoming task. The intuition behind this is that if a task is still processing when a new one comes in, it's likely to be a "big" task, and it's worth letting small tasks ahead in line. Reduces mean response time.
Of course, this could all be moot anyway because context switches are certainly not cheap for software engineers. I just thought I should mention it for nuance. Rules of thumb only get you so far, at some point you have to more accurately model the situation and simulate.
This could be a reasonable assumption if you're very good at adjusting scope, but I would be skeptical unless you have measured it and have the numbers to prove that in practise, tasks indeed turn out to be roughly the same size (±20 % or whatever -- I don't know the exact threshold.)
Why this matters is that under cheap context switches and even modest task size variability, it's -- surprisingly -- a performant queuing policy to pre-empt the currently processing task to handle every incoming task. The intuition behind this is that if a task is still processing when a new one comes in, it's likely to be a "big" task, and it's worth letting small tasks ahead in line. Reduces mean response time.
Of course, this could all be moot anyway because context switches are certainly not cheap for software engineers. I just thought I should mention it for nuance. Rules of thumb only get you so far, at some point you have to more accurately model the situation and simulate.